As general large models draw near, who will foot the bill for the world models?

"We didn't write specific programs for these actions. When it sees the paint palette for cleaning brushes nearby, it walks over to wash the brush on its own and then continues painting." On October 22, Wang Xingxing, founder and chairman of Unitree Robotics (688836.SH), introduced an internal test at the 2026 World Physical AI Conference, where OpenAI's GPT-6 was allowed to directly control a robot to paint without dedicated training.

Wang Xingxing reflected that people's belief in AI is actually far from sufficient, including his own. On November 4, he stated again at the 2026 World Internet Conference Wuzhen Summit that the world's top multimodal large models have already performed extremely well in task planning, and next year will most likely be the critical inflection point for the robotics industry and robotics models.

The multimodal large models Wang Xingxing refers to mainly include general large models such as GPT, Gemini, and Claude. This type of model was originally a large language model designed to predict the next word based on preceding text, and has gradually added image and video processing capabilities since then, with training data primarily sourced from the internet.

GPT-6 Astra, released by OpenAI in September this year, extends the capabilities of such models from task planning to direct robot control.

In a paper published on September 21 by the RoboDojo robot evaluation platform team and researchers from institutions including the University of Hong Kong and Tsinghua University, it was pointed out that without fine-tuning for specific tasks, Astra achieved an average success rate of 22.48% in RoboDojo simulation tasks, while the previous generation GPT-5.5 only reached 0.88%. This result once outperformed all public dedicated robotics models on the RoboDojo leaderboard.

The emergence of high-performance general large models such as GPT-6 Astra has to some extent raised doubts about whether the World Model route, which specifically enables AI to understand the physical world, has been falsified. In other words, since general large models can also control robots, with the continuous improvement of model capabilities, it seems unnecessary to develop dedicated world models.

Is that really the case?

What Can General Large Models Replace

The world model originally refers to a model that can predict the next state of the world based on the current state and a given action, for example, predicting whether a water cup will fall off the table after being pushed by a robotic arm. In the past two years, world models have gained widespread attention, and their connotation has been constantly evolving with the advancement of AI.

As early as 1943, British psychologist Kenneth Craik proposed that there exists a small-scale model of the real world in the human brain, which allows people to predict outcomes before taking actions.

From the late 1980s to the early 1990s, this line of thinking entered neural network research. Researchers David Ha and Jürgen Schmidhuber published the paper "World Models" in 2018, which enabled AI to first learn the changing rules of the environment and then train within the learned model. Since then, the name "world model" has gradually become popular in the deep learning field.

In the following years, world models were mainly a research direction in academia. Turing Award winner Yann LeCun published a paper in 2022 advocating for building machine intelligence with world models as the core. After large language models became the mainstream of the industry, he still publicly criticized such models for lacking an understanding of the physical world.

Since 2024, the popularity of world models has risen significantly, as video generation, robotics, and autonomous driving all require such capabilities.

In February 2024, OpenAI released the video generation model Sora, with its technical report titled "Video Generation Models as World Simulators". In September of the same year, World Labs, founded by Chinese-American AI scientist Li Feifei, made its public debut with $230 million in financing, with the goal of enabling AI to understand the 3D physical world.

In January 2025, NVIDIA released the world model platform Cosmos to generate training data for robotics and autonomous driving. Jensen Huang, founder and CEO of NVIDIA, stated that the "ChatGPT moment" in the robotics industry is approaching.

The world model track was then quickly sought after by capital. After leaving Meta at the end of 2025, Yann LeCun founded the world model company AMI Labs, which announced a $1.03 billion financing in March this year. The company's proposition is: true intelligence does not start from language, but from the physical world. In September, AMD announced its plan to acquire World Labs for approximately $8.2 billion.

Statistics from market research institution VCinno show that there were 23 financing events in the physical AI and world model sectors in 2025, and more than 100 such events have been recorded so far in 2026.

World Labs stated in an article published in June this year that the world model is one of the most ambiguous terms in the current AI field, and divided it into three categories according to output results: the first is generating visual content for human viewing, the second is simulating physical states, and the third is determining the next action.

General large models have now dabbled in all three types of work.

In terms of generating visual content, video generation models launched by general large model companies already have this capability. However, Lin Meibing, chief analyst of ICTIME, told the Economic Observer that realistic visuals do not equal the model's understanding of physical laws. Strictly speaking, video generation models that do not accept action inputs cannot be regarded as world models.

In terms of simulating physical states, general large models achieve this by calling existing tools. A research report released by Zheshang Securities on October 21, 2026 pointed out that Astra can already call the 3D software Blender to reconstruct real spaces, and Google's Gemini 4 Argon can adjust parameters such as friction before the physics engine calculates contact and motion. This means general large models do not necessarily need to master physical laws on their own.

In terms of determining the next action, general large models have seen the most significant capability improvement recently. A simulation test released by NVIDIA's Seattle Robotics Lab in October showed that among 120 grasping and placing tasks, Astra, which received no dedicated fine-tuning for robotics, achieved a success rate of 66.3%, the robotics model developed by NVIDIA based on the world model Cosmos reached 36.8%, and π0.5, a model trained specifically for robotics by the US company PhysicalIntelligence, only achieved 28%.

However, in terms of execution speed, the ranking of the three models is exactly reversed: in the above test, when Astra outputs actions directly, a single call takes an average of 18.56 seconds; NVIDIA's model takes about 0.71 seconds; and π0.5 only takes about 47 milliseconds.

Robot control has very high requirements for speed. Yao Maoqing, co-president of Agibot Robotics and president of the Embodied Business System, previously told the Economic Observer that Agibot's control system is divided into three layers: the bottom layer communicates with the motor 1000 times per second; the middle layer is responsible for motion planning 10 to 20 times per second; and the top layer is responsible for task scheduling. This means for robots, the speed of general large models cannot keep up with the lower two layers, and they are currently more suitable to be deployed at the top layer.

"The production line cannot afford to wait for the time a large model takes to think one step," a product manager at a leading domestic robotics company told the Economic Observer. In terms of specific implementation, the company allows large models to understand tasks and break down steps, while specific actions are still processed by self-trained motion models.

In addition to speed, operation accuracy is also a shortcoming of general large models. On November 5, Wang Xingxing stated during the 9th China International Import Expo that the biggest challenge at present is not the insufficient understanding ability of the model, but the relatively large deviation of "the last few millimeters" when executing in the physical world.

In a real-machine test published by the independent evaluation team Robocurve in September this year, Astra succeeded 19 times out of 20 when putting a block into a bowl, but only succeeded 2 times out of 20 when inserting a precision puzzle piece.

The shortcoming in accuracy is related to training data. "Video learning can let general large models know that a person picks up a cup, but it will not tell it how the robotic arm should move," said the aforementioned product manager. The movement of the robotic arm relies on real-machine data, that is, data collected when the robot operates in practice.

Huang Xiaohuang, chairman of 3D design software company Coohom (00068.HK), stated at the Wuzhen Summit on November 3 that the reason why large language models are the most powerful is that the internet has massive amounts of text data that can be crawled for free, but data from the physical world can only be collected by humans or robots, and the cost remains very high.

Another shortcoming of general large models is predicting the consequences of actions in actual operations. A paper published on October 20 by researchers from institutions including Nanjing University, Nanyang Technological University, and Peking University pointed out that there is a clear gap between Astra's understanding of tasks and its actual execution, with problems concentrated in predicting action consequences, accuracy, and complex motion coordination.

And predicting consequences is exactly the main function of world models, and general large model companies themselves are also developing world models.

On October 27, Google DeepMind released a research preview of the world model Genie4, where the general large model Gemini is responsible for breaking down goals and adjusting plans, and Genie4 is responsible for simulating subsequent scenarios. The two models worked together to simulate a humanoid robot playing checkers.

At least for now, predicting consequences still needs to be completed by world models.

Pure World Model Companies Face "Mounting Pressure"

The ability to predict consequences has its value, but there do not seem to be many buyers willing to pay specifically for world models.

Trial and error in the physical world is different from that in the digital world: if a large model writes wrong code, it can be modified and run again; if a robotic arm knocks over a water cup, the error cannot be undone. The higher the cost of trial and error, the greater the value of advance prediction.

However, this capability does not necessarily have to be provided by dedicated world model companies: most robotics companies and car manufacturers develop it on their own or use open-source models; chip companies open up their models to drive demand for computing power.

A common practice in the robotics industry is to integrate world models into self-trained motion models. Guo Yandong, founder and CEO of robotics company Robosquare, once gave an example that the sequential steps of making tea rely on language models, while predictions such as that a water cup near the edge of the table may fall off rely on world models. Without integrating world models, many tasks are completely impossible to complete.

Currently, robotics companies mainly use world models for two things: first, to deduce possible impacts before executing actions, and second, to generate simulation data. In both use cases, the world model is only part of the full model system of the robotics company.

The aforementioned product manager said that there are already free world models on the market, and the company prefers to invest funds in data and model fine-tuning rather than purchasing dedicated world models.

One of the attractions of world models is the use of low-cost video data to reduce the collection volume of real-machine data, but after comprehensive calculation, the cost is not necessarily lower.

The aforementioned product manager introduced that one method is to first train the world model with internet videos, which requires millions of hours of data, and then fine-tune it with tens to hundreds of hours of high-quality robot data; the other method is to train the motion model directly, which requires thousands of hours of data, most of which is real-machine data with a cost of about 300 to 1000 yuan per hour. The former has a low unit price but large usage volume, while the latter has a high unit price but small usage volume, and the total cost of the two methods may be similar.

Compared with video data, robotics companies pay more attention to real-machine data. Mo Lei, vice president of Robosquare, said in a recent interview with the Economic Observer that real-machine data is more critical for the subsequent performance improvement of models, "the diversity of data is more important than the volume of data."

The situation of car manufacturers is similar: leading enterprises in the intelligent vehicle field mostly develop in-house or develop based on open-source models.

XPeng Motors (09868.HK) released its world model X-World in April this year, built on Alibaba (09988.HK)'s open-source video generation model WAN2.2; the new intelligent driving version launched in September made the world model X-Foresight deployed on vehicles for the first time, which can predict what will happen in the next 6 seconds.

Baidu Group (09888.HK) publicly unveiled its world model ApolloWorld on November 3, which is used for the testing and verification of Apollo Go.

Wu Peidong, managing director of VC Partner, told the Economic Observer that open-source models and tools shorten the starting time, leading car manufacturers will still develop core technologies in-house, and second and third-tier enterprises benefit more significantly. In his view, the weight of intelligent driving in car purchase decisions has risen, but as an independent paid item, consumers' willingness to pay is still very weak. They are more willing to pay for reliable usage and liability guarantee after accidents, and are less willing to pay for "the ability to think", which also leads to insufficient motivation for car manufacturers to pay for external models.

Apart from robotics companies and car manufacturers, chip companies' focus on world models is mainly driven by the demand for computing power.

NVIDIA opened up its world model Cosmos to the industry. Relevant staff told the Economic Observer that opening up these models can help the industry get basic models to a "60-point" level, and car manufacturers can then use a small amount of computing power and their own road test data for fine-tuning. The whole industry's use of these open models will eventually translate into continuous demand for cloud training and simulation computing power.

AMD chose to acquire WorldLabs directly, and the transaction is expected to be completed before the end of the year. Lisa Su, chairwoman and CEO of AMD, stated at the earnings meeting on November 3 that this acquisition will allow AMD to gain deeper insights into the evolution of cutting-edge models. In Lin Meibing's view, what AMD values most is the model team. After integrating the team into its system, the chip company can grasp the requirements of next-generation models for video memory, bandwidth, and latency at an earlier stage.

Robotics companies and car manufacturers rarely pay specifically for external world models, while chip companies open up their own world models, which puts commercialization pressure on world model companies in the market. In addition, the investment required to train world models is also quite large.

Zhu Zheng, co-founder and chief scientist of domestic world model company SuperVision, stated publicly in September that if the training data needs to reach tens of millions or even hundreds of millions of hours, "it may cost hundreds of billions of yuan to train a world model that is truly applicable to general scenarios". Lin Meibing also believes that world models consume a huge amount of computing power, and the computing power scale of startups can hardly compete with that of general large model companies.

In fact, general large model companies are also accelerating their entry into this field. OpenAI announced in March this year that it would stop the consumer application and API services of Sora, and the Sora research team shifted to world simulation to promote the development of robotics.

Sam Altman, CEO of OpenAI, recently stated publicly that OpenAI will manufacture humanoid robots on its own and is expected to release a demonstration prototype in 2027. The reporter also learned during interviews that self-developed hardware can help large model companies obtain high-quality training data more efficiently.

In this context, Zhu Zheng believes that there is no major threshold for general large model companies to shift to world models, and the time window left for pure world model companies is not very long. "We must deliver on the public's expectations within one to two years. Looking ahead to the next five years, the biggest risk or concern I have about world models is whether the commercialization closed loop can be achieved at all," Zhu Zheng said.

This article is from the WeChat official account "Economic Observer", author: Zheng Chenye, authorized for release by 36Kr.